Change bits per channel

[strarsis]

Is there a feature in sharp for changing bits per channel (e.g. from 32bit or 16bit to 8bit)
before processing/saving the loaded image?

[lovell]

At the end of the processing pipeline, before generating the output image, 16-bit pixel values are cast to 8-bit - see

sharp/src/pipeline.cc

Lines 675 to 677 in 7ada9db

	if (sharp::Is16Bit(image.interpretation())) {
	image = image.cast(VIPS_FORMAT_USHORT);
	}

When you say 32-bits per channel, do you mean an 8-bit per pixel image with 4 channels? If so, you may be looking for quantisation, which is not supported.

[jdalton]

I ran into what I think is a related issue with compatibility with pngjs for

As input any color type is accepted (grayscale, rgb, palette, grayscale with alpha, rgb with alpha) but 8 bit per sample (channel) is the only supported bit depth. Interlaced mode is not supported.

Is there was a way to output a compatible png?

[lovell]

Hello @jdalton, the following code seems to produce PNG output images from sharp that can be parsed by pngjs, at least when testing locally with a few different formats from the fixtures.

const assert = require('assert');
const pngjs = require('pngjs');
const sharp = require('sharp');

const input = '5_webp_a.webp';
const width = 100;
const height = 100;
let channels = 0;

const pngParser = new pngjs.PNG();

pngParser
  .on('parsed', (data) => {
    assert.strictEqual(data.length, width * height * channels);
  })
  .on('error', (err) => {
    assert.ifError(err);
  });

const transcoder = sharp(input)
  .resize(width, height)
  .png();

transcoder.on('info', (info) => {
  channels = info.channels;
});

transcoder.pipe(pngParser);

Perhaps a particularly tricky or complex input image is causing problems?

[jdalton]

Perhaps a particularly tricky or complex input image is causing problems?

Ah yeah. Turns out it was a OptiPNG optimization causing the issue.

[lovell]

@strarsis has your original question been answered?

[strarsis]

This was meant to be used to solve this issue: peterbraden/node-opencv#436
So sharp should remove alpha-transparency completely,
including the channel itself so opencv can handle the matching.

[lovell]

@strarsis If OpenCV's matchTemplate feature doesn't support an alpha-transparency channel then I'd suggest using flatten() to remove it.

[strarsis]

Thanks, this works. And for the other way around
(make image 32bit (with alphatransparency)), #554 (comment) can be used.